We start from an animation of the cloud cover predicted by ICON over the Eclipse path.

As we can see, most of the path over the Atlantic will be covered by clouds. Only a small portion of Greenland seems to be cloud free. Over Spain the situation seems to be more favorable, however there are few uncertainties related to some thunderstorms forming in the afternoon.

Here is a map that does NOT represent cloud cover directly, but rather the probability (based on ensemble models) of having clear skies (less than 40% cloud cover, which is already pretty generous).

As we can see, there are a few spots in Spain where the probability of having clear skies at this time is quite low (less than 30%). The area with clear skies in the South-West does not unfortunately coincide with the totality path.

Here are the complete galleries with ICON-EU, ICON and ICON-EU-EPS.

What do I mean by that? Let’s start with a video I made back in 2024 of the total eclipse in the United States. It shows an animation of the eclipse shadow moving across the US and its effect on surface weather, as measured every 10 minutes by the ASOS weather network.

Each wind barb shows wind direction and intensity, the numbers show the actual measured temperature, and the colored scatter points show how temperature varies over time, making it easy to spot even small deviations.

As we can see, the umbra passing over the domain causes a local drop in temperature and a decrease in wind speed, both of which recover once the eclipse has moved on. In fact, many studies have shown that the darkening not only causes air cooling but also a decrease in surface wind speeds (as happening over night) and can eventually affect cloud formation1.

Because many weather models don’t account for the reduced radiation caused by an eclipse, they tend to underestimate its effects on surface temperature, wind, and other variables. That’s why scientists have been working to incorporate eclipse effects into forecast models; a parametrization that’s now present in some global models, including ICON.

A few weeks ago I received an email from DWD (the German Weather Service) saying they had finally added a solar eclipse parametrization to ICON, so I decided to check it out myself. Since we’re looking for a fairly small effect, and the model’s hourly output resolution usually isn’t enough to capture such fast, localized changes, I wasn’t expecting much - but once I downloaded and plotted the clear-sky shortwave radiation, I could see a clear signal.

The first animation shows the clear-sky shortwave radiation flux (essentially solar radiation without the effect of clouds). Watch how it evolves over time: as the terminator (the line between darkness and daylight) moves westward across Europe, you can see a brief “dip” near the eclipse’s center line.

These differences are hard to spot with an untrained eye, so I computed the difference between this radiation field and the previous day’s. That’s not a perfectly clean comparison, since some of the difference simply reflects the changing solar angle as we move toward autumn - but the eclipse signal still stands out clearly as a drop in radiation trailing to the southwest of the eclipse’s center line.

Why to the south-west? I think it has to do something with the sun inclination, projecting a shadow on the west side, and on the hourly averaging of radiation coming out from ICON. Furthermore, notice how the darkening is not entirely visible over Greenland, where the albedo of snow and ice is winning. Arguably it would have been better to use the net short-wave radiation at the top of the atmosphere, but this is unfortunately not availabe in the ICON open data sets of data.

As for the effect on surface weather, it will likely be much smaller this time than what was observed in the US two years ago. In this case, the eclipse’s darkening falls over areas with heavy cloud cover, so its effect will be largely masked at the surface. I couldn’t find any noticeable effect in the surface variables I checked for a few locations in Spain - though the hourly output frequency of the models I have makes it hard to spot any meaningful difference in the first place.

The results speak for themselves. Here is a summary of the entire weather station network’s history from 1853 up to today. Every line in this plot represents the daily maximum temperature recorded across the entire network (effectively the hottest station of the day) for a given calendar year.

Not only did this year mark the earliest exceedance of the 40°C threshold on record (40.1°C in Bad Deutsch-Altenburg on June 29th), but the previous all-time national record (40.5°C in Bad Deutsch-Altenburg on August 8, 2013, during one of the worst heatwaves in recent memory) was completely shattered by a new high of 41.2°C in the same town on August 5th.

Even more striking, 17 of the top 20 all-time Austrian temperature records have occurred in 2026, set consecutively day after day. Here is a selection of those top records:

1. 41.2°C – Bad Deutsch-Altenburg (2026-08-05) [Lower Austria, 168.6m]

2. 41.0°C – Wien Stammersdorf (2026-08-04, 2026-08-05) [Vienna, 190.7m]

3. 40.5°C – Bad Deutsch-Altenburg (2013-08-08) [Lower Austria, 168.6m]

4. 40.4°C – Wien Innere Stadt (2026-08-04) [Vienna, 177.0m]

5. 40.3°C – Neusiedl am See (2013-08-08) [Burgenland, 117.0m]

6. 40.2°C – Krems (2026-08-04) [Lower Austria, 202.0m]

7. 40.2°C – Langenlois (2026-08-04) [Lower Austria, 207.0m]

8. 40.1°C – Wieselburg (2026-08-04) [Lower Austria, 258.7m]

The next plot helps put these extreme events into historical context. I compared the number of stations exceeding fixed thresholds (35°C, 37°C, 39°C, and 40°C) across different years. Because the total number of active weather stations in the network has grown steadily over time, I normalized the data by network size to display relative percentages rather than raw counts.

As the plot shows, the 2013 heatwave was the first time stations ever recorded temperatures reaching 40°C (accounting for just 0.002% of total annual network observations). That same year, observations above 35°C hovered around 1%. This year, that share has jumped to over 2%, while the percentage of stations recording maximum temperatures of 40°C or higher has increased by an order of magnitude (0.02%).

As the data clearly shows, most rivers in Central and Eastern Europe are currently experiencing extreme drought conditions. In some cases, such as the Danube and the Rhine, these extreme conditions have persisted for weeks, if not months.

For example, consider the water level of the Rhine measured at Cologne (Köln). By compiling historical data, we can track these levels back to 1981 (though even older data likely exists in non-digitized archives, but I wasn’t able to find them).

On August 1st, the Rhine’s water level at Cologne reached a staggering historic low of just 68 cm. This dropped below the previous record low of 69 cm, which was set in October 2018. This drastic drop threatens to disrupt key sectors of the German economy (such as the chemical and steel industries) which rely heavily on river transport.

Regarding the river’s discharge (essentially the flux of water passing through), historical records dating back to 1816 indicate that while the August values aren’t the absolute lowest ever recorded (those typically occur in winter) they are extremely close, setting a new record low for the month of August.

Similarly historic conditions have been recorded on the Danube at Budapest. These low water levels have had severe consequences for energy infrastructure. In Hungary, the Paks nuclear power plant (which supplies nearly half of the country's electricity) was forced into an unprecedented shutdown of reactor units because the river level dropped 28 cm below its 2018 record low. The water level fell so much that the intake pumps could no longer draw sufficient water to safely cool the plant. Romania's Cernavoda nuclear plant has faced similar drought-related shutdowns.

There is little need to explain the root cause: a profound lack of precipitation. The following map displays the number of precipitation days for June and July. While not a direct measure of precipitation anomaly, it remains highly representative of the drought’s scale.

Most of Western and Southern Europe has seen few, if any, days of precipitation over the past two months. While an arid summer is typical for places like Algiers, it is entirely anomalous for cities like Paris.

Among these events, the wildfire that developed near Bordeaux, France, was probably the most intense. The smoke plume associated with the fire was clearly visible from space for several days, as shown in the image below, which I created by combining both high- and low-resolution channels from the EUMETSAT MTG-FCI instrument.

I chose this particular time because the satellite image appears to show an overshooting top embedded within the plume. Together with evidence from weather radar and ground photographs, this suggests that the fire was intense enough to generate a pyrocumulus cloud—a type of convective cloud that forms above intense heat sources such as large wildfires or volcanic eruptions.

I also found a photograph on Instagram taken from an aircraft in the area, which appears to capture the same cloud and its distinctive overshooting top.

The impact of the fire on the ground is clearly visible in the before-and-after comparison below. Unlike the previous image, this comparison uses Sentinel-2 satellite data, which provides a much higher spatial resolution than the EUMETSAT FCI instrument.

Here, Sentinel-2 imagery has been processed using a custom composite designed to highlight both active fires and recently burned areas.

Overall, the events of the last few weeks have burned an exceptional amount of land. According to data from EFFIS, this is already France’s worst wildfire season in at least 20 years, and by a substantial margin.

So far, approximately 91,000 hectares have burned, the highest total ever recorded by week 30. This is 96% higher than the previous week-30 record, which was set in 2022. For comparison, the entire 2022 wildfire season—previously France’s worst on record—burned around 66,000 hectares.

The situation in Spain is not yet record-breaking, but it is still highly concerning. Current EFFIS data indicate that 2026 ranks as the second-highest week-30 total on record, behind only 2022, when approximately 227,845 hectares had burned by the same point in the year. The 2026 season has already reached about 91% of that pace.

Although the ignition of individual wildfires cannot automatically be attributed to climate change, the exceptional environmental conditions experienced this summer have almost certainly contributed to their intensity and rapid spread.

The persistent combination of below-average rainfall, unusually high temperatures, and intense evapotranspiration has severely depleted soil moisture across much of Western Europe.

The map below shows the Soil Moisture Index Anomaly, which estimates the amount of water stored within the soil and compares it with the long-term climatological average. The anomalies are expressed in standard deviations, making it easier to assess just how unusual the current conditions are.

Across much of Western Europe—including France, southern Germany, and parts of the United Kingdom—soils are exceptionally dry. In some locations, the anomalies exceed six standard deviations below normal, highlighting the extraordinary nature of the ongoing drought.

These extremely dry conditions are also clearly reflected in the vegetation response, as discussed in my previous post.

This is the same view but with the Sandwich product, which enhances the cloud tops associated with deep convection (red shading) and shows some overshooting tops.

When the night sets in, some wave-like patterns can be clearly seen appearing in the upper-left quadrant. These are caused by the outflow boundary associated with the now dying thunderstorms moving over the ocean. In other words, the cold air “dumped” by the thunderstorm’s precipitation flux propagates away from it, bringing strong wind gusts that creates wave-like perturbation that can be seen once they encounter a cloud layer.

Here are the animations in (hopefully) original resolution on Vimeo.

"You cannot connect a single heatwave to global warming; it’s just an individual weather event, not the climate!"

This sentence, which once served as a cautionary principle within the scientific community, has somehow morphed into a standard mantra for climate skeptics worldwide. Today, I want to show you how, in the context of recent warming trends, single heatwaves can indeed be meaningfully linked to global warming.

First and foremost, we need a reliable way to understand how local temperatures have shifted over time. Imagine gathering all daily maximum temperatures recorded at a specific weather station over decades—taking, for instance, the Lindenberg meteorological observatory near Berlin. By plotting these values, we construct a probability distribution that illustrates which temperatures are most common.

In the first figure, the blue curve represents daily maximum temperatures from an arbitrary past baseline (1940–1950)1, while the red curve shows the same distribution for the recent decade (2016–2026). To keep the comparison fair while the current season is still underway, both datasets focus strictly on the period between May 15 and July 19.

Notice how the red curve has shifted to the right. This shift reflects a clear increase in mean temperature, perfectly consistent with global warming theory. For example, comparing the two curves at 24°C reveals a 1.3-fold increase in probability—meaning a 24°C day, which was already well above Lindenberg’s mid-century average, occurs far more frequently today. Conversely, an event with that same likelihood in the older climate would have been roughly 4 degrees cooler.

The crucial bit, however, lies further down the curve—in the tail, where extreme events live. Focusing on the upper-right inset, which zooms in on hot extremes, we can set a threshold like 39°C. In the mid-twentieth century, temperatures this high were virtually impossible2; the theoretical distribution shows a non-zero probability only because of how the curve was fitted, not because such days actually occurred. In recent years, however, temperatures near 39°C have been recorded multiple times just a few years apart. This drives an estimated 1,800-fold3 increase in probability.

There is an important nuance here: even if the absolute change in probability looks small visually, dividing by a baseline probability that was near zero causes the relative ratio to skyrocket4. That sensitivity in the tail is precisely why extreme events show such dramatic relative increases.

This pattern holds across other regions as well. Looking at Paris (Montsouris), the overall distribution shape differs slightly, and the 40°C threshold shows a more modest 16-fold increase. This lower multiplier occurs simply because high-heat events were already somewhat more frequent in Paris during the baseline period.

To expand this analysis across broader regions, we can transition from individual weather stations to reanalysis datasets. While reanalysis carries its own caveats5, it provides spatial consistency across time. For a broader view of Western Europe, we can subtract the long-term mean (climatology) and divide by the standard deviation, converting raw values into standardized temperature anomalies.

In the accompanying animation covering May 15 to July 15 from 1951 through 2026, we compare yearly standardized anomaly distributions against two baselines (1951–1980 and 1991–2020). The contrast is stark: recent years show an unprecedented frequency of temperatures exceeding 4 standard deviations (+4 sigmas).

The lower panel tracks this trend by plotting the annual probability of exceeding the 1951–1980 99th percentile, highlighting a sharp acceleration that takes off around the turn of the century.

Has this acceleration unfolded uniformly across Europe? Not quite—extreme heat has intensified faster in some areas than in others. To visualize this spatial variation, we fit a Generalized Pareto Distribution (a model better suited for extreme tail behavior than a standard Gaussian) to reanalysis temperatures across the continent. By comparing the likelihood of exceeding the local 99th percentile between 1951–1980 and 2016–2026, we map the relative shift in extreme heat risk.

The results highlight Spain, Southern France, and Italy as particularly vulnerable, which aligns with expectation. More alarming, however, is that virtually all of Europe has experienced at least a two-fold increase in the likelihood of 99th-percentile extreme heat events over this timeframe.

Below is a map of the maximum temperatures compiled from the Icelandic Meteorological Office (IMO), which includes data from both official weather stations and the non-official sensors used to monitor road conditions.

While the western half of the country remained relatively mild or outright chilly, the East Fjords were remarkably warm. The four hottest stations all clustered in this eastern region:

• Seyðisfjörður, 27.4°C

• Skjaldþingsstaðir, 26.9°C

• Seyðisfjörður - Vestdalur, 26.8°C

• Neskaupstaður 26.7, °C

Just how unusual are these temperatures? Unfortunately, most of the stations in this top tier lack long-term historical records. Only the Seyðisfjörður - Vestdalur station has a slightly longer dataset, stretching back to 1995.

Using this data, I mapped the distribution of July maximum temperatures for this specific station to see where this event stands.

Historically, this location sees an average daily maximum of around 12°C in July, making the 27°C reading an incredible outlier. In fact, this temperature sits well above the 99.9th percentile, making it an exceptionally rare event for this area.

Looking at Iceland as a whole, the absolute heat records for July are higher still. The all-time national record for the month belongs to Teigarhorn in the East Fjords, which hit 30.5°C on July 22, 1939. Other notable highs include 29.7°C at Þingvellir (July 30, 2008) and 29.5°C at Hjarðarland (July 14, 2025).

While road-monitoring stations sometimes register even higher numbers, they generally aren’t counted in official climatological archives due to local interference. For instance, a road station at Fagradalsfjall registered a scorching 39.6°C on July 28, 2021—though that particular “heatwave” was courtesy of a nearby volcanic eruption!

While most satellite instruments are optimized for daylight, certain wavelengths (like infrared) allow us to see what’s happening in the dark. One of the most fascinating applications of this is tracking how our global night skies are changing over time, using NASA’s Black Marble product.

A recent study revealed that while the planet is generally getting brighter (up 30% globally), the real story lies in the sharp regional contrasts. The animation below shows the change in nighttime lights over about 10 years (2014-2022): dimming areas are indicated with purple, while brightening areas are indicated with yellow.

The video in full resolution can be found here. This animation was created by me based on the data provided in the original study.

Take a close look at Europe. It is one of the regions contributing the most to global dimming. Notice the heavy purple hues blanketed over France, Spain, and the UK. This isn’t economic decline; it’s efficiency. The dimming is primarily driven by the widespread transition to LED lighting and strict EU-level energy efficiency mandates.

Over the United States and Canada, we see a patchwork. While there are distinct pockets of dimming, the overarching trend is still leaning toward brightening, particularly radiating out from major metropolitan hubs.

The view over Asia tells a completely different story. Here, the map shows sweeping, uninterrupted brightening across almost the entire region, reflecting rapid urbanization and infrastructure expansion over the last decade.

Here a radar animation from the PGUA radar site before it went offline (probably due to damage).

Here is a time series of the stations that were close to the storm: notice how some of the stopped transmitting data at the same time as the radar did.

This animation shows the sea surface temperature anomaly from June 6th to June 29th. The colors illustrate deviations from the long-term average: reds indicate temperatures much higher than normal, while blues show cooler-than-average waters. These data are derived directly from satellite observations.

The impact of early summer European heatwaves is clearly visible here. Notice the progressive spike in sea surface temperature anomalies, which exceed an 8°C deviation on June 29th.

In recent years, the Mediterranean Sea has been losing its role as a natural “dampener” that helps mitigate extreme heatwaves. Because water has a much higher heat capacity than air, it warms up far more slowly, acting as a buffer against sudden temperature spikes. However, this buffer only works if heatwaves remain brief and rare, which, unfortunately, has not been the case lately. Instead, consecutive heatwaves have compounded year after year, leaving the Mediterranean little to no time to recover. As a result, like most of the world’s oceans, it has begun trapping heat deep beneath the surface.

The consequences are clearly visible in this final plot, which visualizes heat accumulation by tracking a specific milestone date. I averaged the sea surface temperature across a zone encompassing the central Mediterranean to find the exact day each year when temperatures first crossed the 26.5°C threshold. Why this specific number? It’s an arbitrary threshold that typically represents the absolute peak of an intense summer season.

As the plot highlights, this extreme threshold used to be crossed around mid-August prior to the 2000s. Only during the infamous summer of 2003 did we see it breached as early as mid-July. Yet, even 2003 pales in comparison to recent years: in 2026, the threshold was shattered before July (on June 28th), following a July 2nd breach in 2025. The long-term trend line indicates that this milestone is shifting roughly 10 days earlier per decade. If this continues, by 2040, the Mediterranean could hit this peak summer temperature by mid June.

Unlike the atmosphere, oceans hold onto their energy. Once heat penetrates the deeper layers, cooling them down takes an immense amount of time and effort, and right now, there is absolutely no sign of a reversal.

Here are the raw images at the original resolution (click to enlarge).

Notice how the average color over land transitions from green to light brown, indicating dry soils and less active vegetation. This data is often used to compute indices (such as the Normalized Difference Vegetation Health Index, or NDVI) that allow scientists to track vegetation health over time.

I tried to compute the average of NDVI over France and compare it to the Interquartile Range (2014-2025). Although this product is only updated every 10 days we can indeed see a sharp decrease in June, much anticipated with respect to the climatology.

The exceptional dryness of the soils can be linked to a lack of precipitation during June, which is the same period where the largest temperature anomalies occurred. The map below shows the precipitation anomalies across Europe for the month of June, expressed as a percentage rather than in millimeters. Values around 0% (light green to light yellow) indicate normal conditions, while an anomaly of 100% represents double the expected amount of precipitation.

Many areas in Spain and France, as well as Southern Italy, were characterized by anomalies as low as -80% in June. While not necessarily unprecedented, this deficit is significant enough to cause vegetation stress and drier-than-average soils. Conversely, a few “wet hotspots” are visible, particularly over the UK and Scandinavia.

Luckily, Europe has a deep meteorological memory. Long before computers, dedicated observers were meticulously logging the weather by hand. Famous centennial time series include:

• Uppsala Observatory, Sweden (started in 1722)

• Bologna Observatory, Italy (started in 1723)

• Padova, Italy (started in 1725)

• Prague-Clementinum, Czech Republic (started in 1775)

By digitizing these paper archives and “homogenizing” the data (adjusting for historical changes in thermometer technology and station locations) scientists have created unbroken records spanning centuries. And no, these data were not “corrected to enhance the warming”, no matter what you read online.

One incredible example is the Paris Montsouris station. Established in 1872 and still active today, it has a virtually gap-free record. This dataset is incredibly useful and allows us to contextually compare today’s extreme heatwaves directly with the past.

Take a look at the absolute maximum temperatures recorded every June from 1872 to the present day.

The data reveals three striking trends. First of all, the maximum temperature recorded 2 days ago, on the 24th of June (40.6°C), was the hottest ever measured in June. Second, this didn’t just break the previous June record, it obliterated it. The former record (37.6°C in 1947) stood for nearly 80 years, and the new record beats it by a massive 3°C. Third, looking at the right side of the plot, it's impossible to ignore how almost all of the highest peaks occur in recent years..

In fact, if we compile a leaderboard of the top 10 hottest June days in Paris over the last 154 years, the footprint of climate change becomes starkly clear. Eight out of the top ten days occurred after 2010, and an astonishing six of those days belong to this year alone!

Even more concerning? The record of this year (so far) is only second in the top 10 of the hottest days ever recorded in this time series. The highest value of 42.6°C was recorded on 25 July 2019 in what is usually the hottest time of the year.

Data courtesy of the MeteoFrance archive.

More precisely, cold pools form when rain falls into the air below the storm, which is typically quite moist. As some of the rain evaporates, it cools the surrounding environment through latent heat consumption, while the remaining liquid rainfall survives all the way to the ground, keeping the surface air chilled.

However, data from the weather station at Belle-Île (an island off the western coast of France) tells a completely different story. Looking at the following plot, we can see that the arrival of the thunderstorm’s precipitation this morning was actually marked by strong wind gusts (up to 100 km/h!) and a sudden rise in temperature, spiking from 27°C to around 35°C.

Data from a neighboring weather station nearby shows that this temperature spike was accompanied by a dropping dew point, signaling the sudden presence of extremely dry air, with relative humidity crashing from 70% down to just 27%.

The radar data from that time shows the thunderstorm passing over the island.

So… what exactly is happening here?

As mentioned earlier, normal cold pools require moist air beneath the storm. In this case, however, the air below the thunderstorm was exceptionally dry due to an ongoing heatwave that had pushed maximum temperatures above 40°C across most of Western France. Because the sub-cloud layer was so dry, the falling rain completely evaporated before ever reaching the ground.

This rapid evaporation created a pocket of incredibly dense air. Gravity took over, causing this heavy air pocket to plunge rapidly toward the surface into layers of increasing atmospheric pressure. With no liquid water left to continue evaporating and cooling the air, the air began to warm up rapidly due to compression. This process of compressional heating is highly similar to how Foehn winds warm up as they cascade down the Alps.

An atmospheric sounding from Bordeaux, located relatively close to the island, perfectly captures this setup. It reveals a layer of intensely dry air trapped between the surface and 700 hPa. Notice also how the atmosphere is highly unstable, yet the convection is tightly capped by strong surface inhibition. This is a classic meteorological “loaded gun” scenario, where the right trigger can instantly unleash a violent weather event.

Documented cases of these spectacular “heat bursts” are relatively rare, which makes analyzing anomalies like the Belle-Île event all the more fascinating.

High resolution video.

The animation at the top of this article depicts an incident from May 22, 2026. A Cathay Pacific Airbus A350 (Registration: B-LRV, Flight: CX156) flying over the Philippines encountered severe turbulence, resulting in injuries to 10 people.

The video (condensing about 10 minutes of real-time flight) overlays a small portion of the flight track (with an enhanced vertical scale) onto the satellite imagery recorded at the exact moment of the incident. While the initial onset of turbulence is subtle, you can clearly recognize the ensuing drop in altitude. This was most likely a deliberate maneuver commanded by the pilots to escape what appears to be convective-induced turbulence.

The chart below breaks down the aircraft’s altitude, vertical speed, and ground speed during the encounter, captured via ADS-B data.

The data illustrates the likely sequence of events.

1. An initial rapid pitch-up (a sharp increase in altitude and vertical speed) likely caused by a powerful convective updraft.

2. A subsequent descent at a rate of roughly 2,500 feet per minute, representing the pilots’ proactive response to exit the turbulent zone.

3. Just a few minutes later, the aircraft safely climbed back to its original cruise level around 40,000 feet.

To truly understand what happened, we have to look at a wider spatial and temporal view. As Flight CX156 approached the island of Marinduque, an isolated thunderstorm exploded into existence almost out of nowhere. In the satellite imagery below, this is indicated by the deep red areas. These colors signify incredibly low cloud-top temperatures, a definitive fingerprint of a severe, violent thunderstorm.

Because this satellite data is updated only every 10 minutes (and the 20:40 UTC frame is unfortunately missing for this day, not sure why) perfectly aligning the imagery with the exact second of the flight data is difficult. However, the delayed frames still clearly pinpoint the rapid birth of the storm cell.

The sheer velocity at which this thunderstorm materialized is almost certainly the reason the flight crew had no time to avoid the encounter.

Today’s aviation toolkit is undeniably amazing. Flight dispatchers and pilots use sophisticated forecasting models to map out and avoid turbulent regions during pre-flight planning and in real-time. However, when a thunderstorm intensifies this rapidly, it presents a unique hazard. They can mature faster than satellite loops update, and sometimes even faster than an onboard weather radar can fully resolve, leaving flight crews with mere seconds to react.

Among the temperatures that were recorded, I’d like to cite the 35.1°C measured at Kew Gardens (London), the 40.3°C recorded at Mora (Portugal), the 39.5°C recorded at Vinebre (Spain), the 39°C measured at Fitou (France), and lastly the 30.4°C measured at the observatory of De Bilt (Netherlands).

Many of these values were station (or even nationwide) records. However, taking only the station data into consideration makes it hard to draw a complete picture of the heatwave over all of Europe because these records are isolated and span different time periods.

For this reason, I used reanalysis data, which standardizes the observations into a space that is statistically consistent over time, and focused on the countries that were the most affected by the heatwave. This first picture shows the average (over the area shown in the inset) of the daily maximum temperature: the red line shows this year compared to all the previous years (up to 1951), which are represented with black lines.

As you can see, this heatwave was unprecedented in the reanalysis record.

What is also surprising is that this was not a localized event but affected almost the entire continent. This can be seen in this second figure, where I tried to use the same reanalysis data but in a slightly different manner: instead of directly comparing the temperatures of this year to the previous ones, I ranked all these years by focusing on the period from 22 May to 5 June.

The color scale allows us to find the areas where the aforementioned period was the hottest on record by looking for the dark red color: not surprisingly, we find most of Spain, Portugal, France, and the UK. The rest of the continent also experienced warm conditions, although not record-breaking. Only Russia and neighboring countries experienced normal or even cold temperatures.

We can “get rid” of the dependency on the limited period covered by the reanalysis and instead compare the heatwave to a long-term climatology (1991–2020). This way, we can assess how warm the period was with respect to a reference without considering whether a similar event happened before.

The following picture shows the anomaly of the maximum temperature (the average is still computed on the same region depicted in the first figure) day by day from February until June.

As you can see, the event stands out, as it exceeds three standard deviations of the temperature anomaly, making it extremely rare (about a 0.1% probability).

But how exactly are heatwaves evolving in Europe? One naive way to track them is to look only at daily maximum temperatures from reanalysis (the closest model we have to observations). Although one can define many indices to track heatwaves (for example, the number of consecutive days with a maximum temperature above a certain seasonal threshold) in this article, I decided to use something much simpler: the total number of days (per year) with a daily maximum temperature exceeding 30°C. Of course, reanalyses are still models and, as such, cannot perfectly represent local temperature peaks. With this out of the way, let’s have a look at some results.

Let’s start with a picture of the past (an average of the 10-year period between 1965 and 1975). This plot shows the number of days per year with a daily maximum temperature above 30°C, ranging from white (2 days per year) to dark red (120 days per year, which is longer than the 3 months of summer).

From the map, it is immediately apparent that (on average) days with maximum temperatures above 30°C were extremely rare (if not impossible) in Central and Northern Europe, including the UK. While almost all areas south of the Alps experienced about 10 to 45 days, southern Spain used to be the hottest spot, with maxima around 100 days per year.

Fast forward to the 2015-2025 period, and things have changed drastically. First of all, many areas that did not experience temperatures above 30°C at all (e.g., the UK, Northern Germany) now have many hot days (ranging from 2 to even 30).

As a consequence, areas that were already hot before are getting even hotter. Southern Spain is still the hottest area, with local peaks above 120 days around Cordoba. Similar patterns can be found in all Mediterranean countries. In fact, across Southern Europe, the only areas that do not have any days above 30°C can now be found only in the mountains or close to the Atlantic (where, however, the effect of humidity evaporating from the sea may increase the feels-like temperature, even with maxima below 30°C).

Here is an animated gif that makes the comparison even easier.

Are these trends significant? As a matter of fact, yes, almost everywhere. In the final plot I’m showing, I computed the trend of the same variable (days per year exceeding a daily maximum temperature of 30°C) from 1951 all the way to 2025. Areas with non-significant trends are masked and negative trends are not represented in the color scale (for obvious reasons).

From this plot we can indeed see that, for many regions, the “hot get hotter” paradigm, as the strongest trends can be found in areas that were already experiencing extreme temperatures in the older climate.

The heatwave of May 2026 shows that Europe was, and still is, a hotspot of global warming and heatwaves intensification.

Additional plots over the full period

Methods/code

Let’s load the data from Earth Data Hub. Note that, instead of using the already computed daily data, I had to load the hourly version and compute the daily maximum temperature myself. This means that the computing part is going to take quite a long time (few hours in my case). This means using the cache feature is quite smart in case we need to stop and restart later

import xarray as xr
import cartopy.crs as ccrs
import cartopy.feature as cfeature
import matplotlib.pyplot as plt
import numpy as np
from dask.diagnostics import ProgressBar
import matplotlib.patheffects as path_effects
import rioxarray
import matplotlib.dates as mdates
import matplotlib.colors as mcolors

PAT = {your_token}

ds = xr.open_dataset(
    f"asyncwrapper::simplecache::https://edh:{PAT}@data.earthdatahub.destine.eu/era5/reanalysis-era5-land-no-antartica-v0.zarr",
    chunks={},
    storage_options={
        "https": {"client_kwargs": {"trust_env": True}},
        "simplecache": {"cache_storage": "/tmp/edh_cache/era5-land"},
    },
    engine="zarr",
)

ds = ds.rio.write_crs(ccrs.PlateCarree())

lons = (-25, 50)
lats = (35, 72)

# Subset in space over Europe (sticking together two portions)
subset = xr.concat(
    [
        ds[["t2m"]].sel(
            longitude=slice(0, lons[1]),
        ),
        ds[["t2m"]].sel(
            longitude=slice(360 + lons[0], 359.9),
        ),
    ],
    dim="longitude",
).sel(latitude=slice(lats[1], lats[0]))

Next we define the threshold and count all days that have at least 1 timestep with temperature above it.

thresh = 30
exceedance = subset["t2m"] > 273.15 + thresh
exceedance = exceedance.coarsen(valid_time=24, boundary="trim").any()

hot_days = exceedance.groupby("valid_time.year").sum(dim="valid_time")

with ProgressBar():
    hot_days.load()

Note that I haven’t used a .resample().max() because, from my testing, this appeared to be quite slow. The .coarsen().any() operation should produce a similar result and appears to be much faster because the time index does not need to be checked,.

The plotting part is quite easy; I just truncated the colormap to get rid of the extreme colors from the gist_heat colormap.

years = (1995, 2025)
fig, ax = plt.subplots(
    figsize=(13, 13),
    subplot_kw={"projection": ccrs.Mercator()},
)

ax.set_extent([-18, 40, 35, 65], crs=ccrs.PlateCarree())

ax.add_feature(cfeature.LAND.with_scale("10m"), facecolor="#424e59", zorder=0)
ax.add_feature(
    cfeature.COASTLINE.with_scale("10m"), edgecolor="white", zorder=7, linewidth=0.2
)
ax.add_feature(cfeature.OCEAN.with_scale("10m"), facecolor="#2c2f3c", zorder=0)
ax.add_feature(
    cfeature.BORDERS.with_scale("10m"), edgecolor="#58646e", linewidth=1, zorder=5
)

p = hot_days.sel(year=slice(years[0], years[1])).mean(dim='year')['t2m'].plot.contourf(
    ax=ax,
        cmap=colors.LinearSegmentedColormap.from_list(
        "test", plt.get_cmap("gist_heat_r")(np.linspace(0, 0.9, 256))
    ),
    levels=[2, 5, 10, 30, 45, 60, 90, 120],
    transform=ccrs.PlateCarree(),
    extend="max",
    add_colorbar=True,
    cbar_kwargs={
        "shrink": 0.3,
        "label": "",
        "anchor": (-6.1, 0.3),
    },
)

But how can we quantify the magnitude of a heatwave in an area with only a few weather stations? A good proxy is the height of the freezing level, which is the altitude at which the temperature first reaches 0°C when descending from the upper atmosphere toward the surface. This definition is crucial because we don’t want to factor in surface inversions, which often occur, especially in mountain valleys.

Freezing levels can be estimated using data from radiosondes, weather balloons equipped with instruments to measure temperature, humidity, and other atmospheric variables. These kinds of measurements are vital for weather forecasts, as they offer the only way to obtain in-situ (on-site) observations higher up in the atmosphere.

The Swiss Meteorological Office (MeteoSwiss) has been launching radiosondes from Payerne (west of Bern) since the 1950s, meaning we can track the evolution of many indices, including the freezing level, over a long period. Using this entire archive, I computed the climatological average (1981–2010) of the freezing level and compared it to the values we’re observing in the last few days. Have a look at the resulting plot.

Since May 22nd, the freezing level has been hovering above 4,000 meters, with peaks of 4,250 meters reached around May 25th. These are unusual values not just for May (when the freezing level typically hovers around 2,600 meters) but even for August, when the climatological average peaks around 3,600 meters. This means we’re not just months ahead of schedule; we’re already exceeding typical summer values.

Moreover, the current warm spell is so prolonged that it has beaten every month of May in the last 70+ years, as shown in the second plot. Here, I counted the number of days in May per year where the freezing level exceeded 3,900 meters (a value unusual even for summer, as noted before).

In May 2026, there have already been 7 days where this threshold was exceeded. This has never happened before and sets a new record (breaking the previous record set in 2016).

How to reproduce this analysis

The data was downloaded from the IGRA archive for the station SZM00006610. The archive starts in 1949 and is frequently updated up to today. The dataset (loaded into a pandas dataframe df) will look like this.

We can proceed by computing a monthly climatology on a restricted period.

df['month'] = df.index.month

res = (
    df.loc["1981":"2010"]
    .groupby("month")
    .mean()
    .rename(columns={"FRZHGT": "FRZHGT_monthly_mean"})
    .merge(df.reset_index(), left_on="month", right_on="month")
    .set_index("date")
    .sort_index()
)

That’s already everything we need to plot the data.

Let’s first draw the horizontal lines showing the monthly climatological values.

subset = res.loc["2026-05-20":"2026-05-28"]

fig, ax = plt.subplots(figsize=(9, 7))

months = [5, 6, 8]
for i, month in enumerate(months):
    mean_val = res[res["month"] == month]["FRZHGT_monthly_mean"].iloc[0]
    ax.axhline(
        mean_val, color="black", linestyle="-", linewidth=1, alpha=0.5, zorder=1
    )
    rightmost = subset.index.max()
    month_name = calendar.month_name[month]
    ax.text(
        rightmost + pd.Timedelta(days=0.5),
        mean_val,
        f"{month_name}",
        va="center",
        ha="left",
        fontsize=15,
        color="black",
    )

Now let’s plot the recent data using a stem plot.

vals = subset["FRZHGT"].values
cmap = plt.cm.coolwarm
norm = plt.Normalize(vals.min(), vals.max())
colors = cmap(norm(vals))

markerline, stemlines, baseline = ax.stem(
    subset.index,
    subset["FRZHGT"],
)
stemlines.set_linewidth(8)
stemlines.set_colors(colors)
baseline.set_linewidth(0)
markerline.set_markersize(0)
ax.scatter(
    subset.index,
    subset["FRZHGT"],
    c=colors,
    s=350,
    edgecolor="k",
    linewidths=0.5,
    zorder=3,
)

We used a color mapping for both the lines and the markers making up the stem plot.

Some formatting and (optional) adding an image background.

ax.xaxis.set_major_locator(mdates.DayLocator(interval=1))
ax.xaxis.set_major_formatter(mdates.DateFormatter("%-d %b"))

ax.set_ylabel("height [meters]")
ax.set_xlabel("")
ax.set_title("Freezing levels in the Alps: May like August")
ax.grid(False)
ax.set_ylim(subset['FRZHGT'].min() - 250, subset['FRZHGT'].max() + 250)

img = plt.imread('./background.png')  # Update with your image path
ylims = ax.get_ylim()
xlims = ax.get_xlim()
ax.imshow(img, aspect='auto', extent=[xlims[0], xlims[1],
ylims[0], ylims[1]], zorder=0, alpha=0.5)